I think the major issue is that of dtypes. I had imagined a number of passes to add other sorting algorithms and improvements through writing tests. I agree that it doesn't work for virtual arrays. (I think you mean stuff like slices, right?). I guess my brief question is, if people are interested in contributing to numpy/scipy development, are there specific helpful things to do? Perhaps a to do list with tests or a roadmap? Or just wait for the dtypes branch to be merged to the default and then go from there?
I don't think this is complete. The algorithm looks a bit to simple and also it can just reuse timsort from rlib with a little work. Also, as for a sorting algo reimplementation it definitely lacks tests (there is no odd-sized list, no empty list for one)